Ecommerce Product Chatbot: Guide Without Annoying

The problem is choice, not information

A shopper on a catalogue of four thousand items does not need more information. They need the field narrowed to three things they could reasonably buy. That is a different task from search, and it is the task most recommendation bots fail because they are built as a search box with a personality.

The measure of success follows from that. A good product conversation ends with the shopper looking at two or three items and understanding the difference between them. A bad one ends with a list of twelve, which is what they already had.

It also means the bot's value is highest exactly where the catalogue is hardest — variants, sizes, compatibility, technical specifications. On a small, simple catalogue a good filter beats a conversation.

The questions worth asking

Every question costs attention, so each one has to eliminate a large part of the catalogue. Three or four is the practical limit before a shopper leaves.

1. 1The constraint questionThe one attribute that rules out most of the range: size, compatibility, capacity, room dimensions, device model. Ask this first because it does the most work, and it is the question shoppers expect to be asked.
2. 2The use questionWhat it is for, in their words. This is where a conversation beats a filter: 'for a child who is just learning' selects differently from 'for daily professional use', and no filter captures it.
3. 3The trade-off questionWhat matters more — price, durability, speed, appearance. Asked once, it turns a list into a recommendation, and asked as a genuine choice rather than an upsell it does not read as selling.
4. 4The budget question, last and lightlyA range rather than a figure, offered as options rather than an open question. Asked first it reads as qualifying the shopper; asked last it reads as narrowing the shortlist.

Presenting the result

• Offer two or three options, never more. A list of ten is the problem the shopper came with.
• State the difference between them in one line each. 'This one is cheaper; this one handles heavier use; this one is in stock today' is a recommendation. Three identical descriptions are a catalogue.
• Confirm availability at the moment of recommending, from live data. Recommending something out of stock is the fastest way to lose the sale and the trust together.
• Include the concrete blocking facts: delivery time, whether it fits, whether it is returnable. These are the questions that otherwise arrive after the recommendation and stall the purchase.
• Link to the product rather than describing it at length. The bot's job is to get them to the right page, not to replace it.
• Make it easy to reject all three and start again, without the bot restarting from question one.

Where a shopping bot becomes annoying

Most of the damage in this category is self-inflicted and comes from a small number of behaviours. They are worth naming because each one is easy to remove.

• Opening unprompted with a question. A shopper reading a product page did not ask for an interview.
• Asking for contact details before giving anything useful. In ecommerce this converts a browsing session into a closed tab.
• Recommending an upgrade when the shopper stated a constraint. If they said the budget is a hundred, do not lead with the two-hundred option.
• Re-asking something already said, which reads as not listening and is the most common abandonment trigger.
• Enthusiasm about every option, which destroys the credibility of the recommendation. A bot that says everything is excellent has recommended nothing.
• Blocking access to a person, especially where an order already exists.

Live data, not indexed data

Product conversations depend on facts that change faster than any indexed page, and getting this wrong is more damaging here than in most sectors because the error is discovered at checkout.

• Stock and variant availability — the single most important live value in ecommerce.
• Current price, including active promotions, which change faster than material is updated.
• Delivery estimates to the shopper's location rather than a generic figure.
• Order status for existing customers, which is a support conversation arriving in a sales channel.
• What is genuinely safe to index: specifications, compatibility rules, materials, care instructions, returns policy and size guidance.

Where to hand over

• An existing order or a delivery problem — this is support, not sales, and routing it to a product conversation is the most common misroute in ecommerce.
• Any request for a discount beyond published promotions.
• A return, refund or warranty claim, which is a decision rather than an information question.
• A complaint of any kind, immediately and without a qualifying question.
• Bulk or business enquiries, where the price and the terms differ from the shop.
• Anything about safety, allergens, medical suitability or regulated claims, which should never be improvised.

What to measure

1. 1Add-to-cart rate from conversationsConversations that ended with an item added, divided by product conversations. This is the number the flow exists to move.
2. 2Conversations per recommendationHow many exchanges before a useful recommendation appears. Rising numbers mean the flow has grown a qualification step it does not need.
3. 3Out-of-stock recommendation rateRecommendations of items unavailable at the time. This should be zero and it will not be unless availability is checked live.
4. 4Return rate on assisted purchasesCompared with unassisted. A rising return rate means the bot is recommending confidently rather than accurately, which is worse than not recommending.
5. 5Support misroute rateOrder problems that landed in the product flow. A high figure means intent detection is running after topic rather than before.

How Vexvon handles product conversations

The catalogue is a first-class knowledge source: a product catalogue with variants and images can be loaded alongside your site crawled to as many as 4,500 pages, PDFs up to 25 MB and bulk entries at up to 200 per request. Entries are typed into thirteen categories including product, catalogue, pricing and campaign, so a promotion can be added and removed without touching the specifications.

Stock and price should not be answered from an index, and do not have to be: your own API can be registered as a tool the agent calls mid-conversation, with protection against unsafe destinations. Where the lookup fails, the defined behaviour is to say it could not be checked rather than to recommend something that may not exist.

Two mechanisms matter for the discovery conversation specifically. Query rebuilding turns a fragment such as 'what about the red one?' into a complete standalone query before retrieval runs, which is exactly the shape product conversations take. And an image sent by the shopper can be matched to a product through a high-confidence shortcut, which shortens the constraint question considerably.

On the support boundary, all channels write to one customer record, so a shopper with an existing order is recognised rather than treated as a new enquiry. A request for an operator raises a notification, an agent's stop character pauses the AI for thirty minutes, and conversation memory of the last fifteen messages travels with the handover.

Frequently asked questions

1. 1What does a product recommendation chatbot actually do?It narrows a catalogue to two or three items the shopper could reasonably buy and explains the difference between them. That is a different task from search, and most bots fail it by behaving like a search box with a personality.
2. 2How many questions should it ask?Three or four at most: the constraint that eliminates most of the range, what it is for, the trade-off that matters, and budget last and lightly. Every question costs attention.
3. 3What makes a shopping bot annoying?Opening unprompted with a question, asking for contact details before helping, recommending an upgrade against a stated constraint, re-asking what was already said and being enthusiastic about every option.
4. 4What must come from live data?Stock and variant availability, current price including promotions, delivery estimates to the shopper's location and order status. Specifications, compatibility and policies can safely be indexed.
5. 5When should a person take over?Existing orders and delivery problems, discount requests, returns and warranty claims, complaints, bulk enquiries and anything touching safety or regulated claims.
6. 6Which metric matters most?Add-to-cart rate from conversations, watched together with the return rate on assisted purchases. A bot that recommends confidently rather than accurately moves the first and quietly ruins the second.

Read twenty abandoned product conversations

Take twenty conversations where a shopper asked about a product and did not buy, and mark the message at which they stopped. In most implementations they cluster on one specific question — usually the second one, usually asked before anything useful was given. Moving that question one step later is often the whole improvement.